Confession: I have always hated the phrase "design sense."
It was 2016, when I was attending business school. For the final project, I spent nearly eight hours creating slides. I gathered massive market data, organized case studies neatly, and was confident in a document that covered everything from 1 to 10.
But the instructor's words flashed through my mind. "You've just organized information; aren't you suffering from analysis paralysis? So, what is the storyline?"
It felt like being hit on the head. My 18 slides were nothing more than a "missing picture book" of neatly arranged data. At midnight, with a churning stomach, I threw two-thirds of the eight-hour work into the trash.
In the remaining time, I kept only what was necessary to answer the "issue (question)," sharpening each slide to contain exactly one message. The next day's presentation. The stripped-down slides brought unprecedented praise and specific feedback.
Stopping the cramming of information for perfection and having the courage to "discard" based on standards. This exhilarating experience became the origin etched into my bones.
……And yet.
Time flows to 2026. AI has evolved, and everyone shouts, "One prompt for finished slides!" I found myself back in the same nightmare as 2016.
"Copy the text." "Paste it into the AI screen." "Ask for 'nice-looking slides.'" "AI kindly produces a cyber-space with rainbow gradients." "Text sizes are mixed up in about 8 varieties, flooding with information." "Desperately writing long correction requests like 'No, not that, the number in the square at the top right of slide 3...'" "When fixed, another layout breaks."
……Why are humans recreating "analysis paralysis and design mudslides" against the latest AI? It's completely absurd. I'm losing my mind.
Letting AI create slides with free-spirited sense means humans will be holding their heads again late at night. What we want for work isn't AI's whimsical art pieces. We want "a proposal that conveys its point in one slide and gets approved tomorrow."
To end this futile struggle, I introduced the Digital Agency Design System (DADS) and a "Red Frame Review Board" that simply surrounds images.
Stop Asking AI for "Sense," Give It the Country's "Blueprint (DESIGN.md)"
Asking AI to "make it look nice" is like asking an architect to "build a nice house." One day it's a log cabin, the next a skyscraper. Blaming AI is wrong; the fault lies with the human who didn't provide a blueprint.
At work, when making UI mocks with AI (Gemini or Claude), I always provide a rulebook called "DESIGN.md" beforehand. Just inserting this minimizes the variance in the resulting screens.
Slides were exactly the same. That's where I noticed the "Digital Agency Design System (DADS)" published by the Japanese government.
"Isn't a national design system stiff and plain?" you might think.
Not at all. As a business document, it is the ultimate example of "beautiful subtraction."
It's clean, unambiguous for anyone, and information goes straight into your head. Best of all, the Digital Agency distributes these guidelines in "Markdown format text," which AI understands most easily.
Hand this Markdown file to Codex or Claude Code and say: "Compose the manuscript into slides following these Digital Agency standards."
Just like that, accidents where AI enthusiastically scatters neon colors are permanently zero.
https://x.com/tetumemo/status/2096924906604958133
5 Text Sizes, 3 Spacing Levels. Rules Give You the "Courage to Strip Away"
What happened after handing over the Digital Agency standards? The "stripping away" I once did manually at midnight is now executed mechanically by AI as a convention.
Font sizes were strictly fixed to these five levels:
- Cover Title: 48px
- Main Heading: 36px
- Subheading: 24px
- Body Text: 20px
- Footnotes/Annotations: 14px
Whimsical ideas like "let's make this 28px to stand out" are absolutely forbidden. Headings grab attention, body text explains reasons, footnotes are checked if needed. This visual guidance is completely fixed across all slides.
Spacing also uses only three types: "8px, 24px, 64px." Stop the bad habit of filling empty space with meaningless illustrations because "the bottom right looks lonely," and let whitespace function as "information separators."
Even more powerful was the "color contrast ratio" check.
My brand color for regular posts is a vivid yellow (#fbd212). When I tried making slides with this beloved color, the design system's inspection program rejected it harshly.
The contrast ratio against a white background was "1.47:1." It fell far short of accessibility standards of "4.5:1 or higher for text" and "3.0:1 or higher for graphics."
"But that's my symbol color...!" I was shocked for a moment, but of course. If participants squint at thin yellow text on a projector or rough display, they just get tired. Yellow was relegated to decorative bands on covers, while dark blue (#155585, contrast ratio 7.87:1) was applied to readable text.
Correcting subjective preferences with objective "numbers." This is the greatest benefit of using AI as a work tool.
The Red Frame Erased the Pain of Writing "Top Right of Slide 3..."
Even with a stable skeleton, a huge irritation lurked in the field: "Giving correction instructions."
Looking at AI-generated slides, the moment you think "emphasize this number more" or "split this footnote into two lines," the tedious work begins.
"On slide 3, under the heading 'Sales 60 Plans' inside the card on the right, the small number written below it..."
……I almost go crazy writing this. The place to fix is right in front of me, so why must humans compose Japanese sentences describing "coordinates"? Worse, AI misreads the desperate explanation and rewrites an unrelated top-left heading. Screaming "That's not it!" at the screen is a hollow waste of time.
To eradicate this pain, I created a browser-based "Red Frame Review Board."
Usage is surprisingly simple. Drop the slide image into the browser, draw a red frame around the area to fix with mouse or finger. Then write just "Make text larger" or "Split into two lines" in the input box and press the button.
Instantly, the correction instruction for AI is generated.
Slide 3 [x: 620, y: 340, w: 200, h: 80] area: Make text larger
Just paste this into Codex or Claude Code. AI sees the image and coordinates, says "Understood," and fixes that spot precisely to the millimeter.
Humans naturally want to instruct like pointing at a colleague's shoulder saying "Fix this here." The essay-writing time to describe locations verbally has vanished from this world thanks to the Red Frame Review Board.
https://x.com/tetumemo/status/2096167671910076617
Even if the Score (Blueprint) Passes, Don't Trust Until You Hear the Performance (Actual File)
But here lay the biggest pitfall.
When building a system to automatically convert HTML slides viewed in browsers into PowerPoint (.pptx) format for direct distribution, I ran an automated inspection program checking for text overflow and color contrast. The result was beautifully "ALL GREEN (All Items Passed)."
"Yes, perfect! Ready for distribution!" But the moment I opened the actual file in Windows PowerPoint, my breath stopped.
……It was spectacularly broken. In five places.
- The third line of the source citation is cut off below the frame.
- The number card heading repeats "Achievement Achievement."
- Table height is insufficient, causing a collision accident with the note below.
- Summary lines overlap with cautionary notes, making them unreadable.
- "Total 60 Plans 60 Plans" mysteriously duplicates in the center of the donut chart.
"You said all items passed?! Why is it overlapping?!"
I identified the cause and deeply reflected. The program was inspecting the "frame height at the time of file generation." However, the moment the actual PowerPoint app opens the file, fonts and line wrapping cause the frames to stretch downward unexpectedly.
In other words, things that didn't overlap in the blueprint stage collided in the rendered reality.
I had experienced the exact same thing when researching AI generation for Reels videos. No matter how perfect the prompt (score), when viewing the actual AI-generated video (performance), fatal bugs like characters eating food through masks occur casually.
Testing the "Score (Blueprint)" and Quality Control (QC) of the "Performance (Actual Object)" must be operated separately.
No matter how proudly AI says "Passed!", we must never remove the checkpoint where humans verify the final actual object with their own eyes. I hurriedly modified the logic to calculate heights after text expansion, incorporated a step to launch PowerPoint in the background, image every slide page, and perform final acceptance with human eyes.
Automation that hasn't passed through muddy verification is just a ticking time bomb in the field.
Where I Ended Up After Burning Tokens Flying Pterosaurs
Right before creating this slide production kit, I was obsessed with the terrifying evolution of OpenAI's latest model "GPT-6 Astra."
Overseas demos showed walking 3D spaces built from floor plans, moving furniture and coffee machines. Unable to contain excitement, I threw an unreasonable request at Codex Astra: "Create a cinematic shooting game flying over ruins on a pterosaur in WebGL style like Panzer Dragoon!"
https://x.com/tetumemo/status/2096195768726216992
https://x.com/tetumemo/status/2096005966789759350
Result: A playable 3D game appeared in the browser in just 20 minutes. Shocking and moving. ……But the cost was hitting the usage limit instantly, burning precious reset tickets and consuming "200%" of usage volume.
After playing the game, sitting in a silent room drinking cold tea, I came to my senses.
"……Am I going to fly pterosaurs every day for work starting tomorrow?"
Absolutely not. What I really want to do in the field tomorrow is quickly create proposal materials that resonate with bosses and clients, and be freed from late-night micro-adjustments.
Don't let burning tokens become the goal itself while dazzled by flashy AI evolution. Borrowing the power of cutting-edge models, keep the "grounded template" of Digital Agency standards at hand, taming AI as a tool to reliably advance daily work.
This is the true realism for us non-engineers to survive the AI era.
Have the Courage to Strip Away, Together with AI
At midnight in 2016, when I threw two-thirds of eight-hour slides into the trash, I was alone. Unsure of the right answer, I trembled pressing delete based only on the intuition that "this won't convey."
But now is different.
- Thinking Role (Human × AI): Prepare manuscripts, sharpen the message (title column) of each slide.
- Organizing Role (Digital Agency Design System × Generation Skills): Instantly assemble stable slides following national conventions for font size, spacing, and color.
- Verifying Role (Red Frame Review Board × Human Eyes): View the actual object, point-and-shoot corrections, guarantee final quality with humans.
The essence of document creation isn't competing in flashy decorations. It's stripping away noise to deliver clear messages to the audience's brain.
The standards for that are already prepared by the Digital Agency. The red frame to communicate fixes is also at hand.
No need to be tossed around by AI's whimsical designs anymore, nor stay up all night tweaking text sizes. I want to proudly hand these tools to my past self, gasping for air in front of the screen.
Full data for the "Digital Agency Design System Compliant Slide Skills Kit (dads-slide-deck-v1)," Red Frame Review Board, and PowerPoint conversion scripts introduced here are explained and distributed in the original article below.
📝 Over 6,000 subscribers! Publishing weekly experiential newsletters focused on AI. Free to read ↓
